--- title: The Communication Bottleneck: A Round-Trip Study of... description: A research study introduces a round-trip protocol to measure how much tree-structured information (like arithmetic expressions) survives when language models... canonical: https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models | daily.dev og:description: A research study introduces a round-trip protocol to measure how much tree-structured information (like arithmetic expressions) survives when language models... og:url: https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf og:image: https://api.daily.dev/og/posts/ZAAGsxLyF.png og:image:alt: The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models **[Apple Machine Learning Research](https://daily.dev/sources/mlapple)** · 2 min read · 0 upvotes · 0 comments ## Summary A research study introduces a round-trip protocol to measure how much tree-structured information (like arithmetic expressions) survives when language models serialize it into natural language and another model tries to recover it. Testing sixteen models pairwise reveals the channel is lossy and asymmetric, with accuracy varying by up to 60.4 points depending on which models generate versus extract; the best pairing reaches 92.9% accuracy. Most failures (73.6%) originate at the generation step, driven by tree structure complexity rather than model family. Fine-tuning on ~3600 examples improves every open-weight model beyond untrained Gemini-3.1-Pro, even in a disjoint domain with new operators, though a gap to frontier models persists. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much does accuracy vary when swapping which language model generates versus extracts structured information in a round-trip serialization task? Accuracy shifts by up to 60.4 percentage points depending on which model generates the word problem and which extracts the arithmetic expression back. The best-performing pair reaches 92.9% accuracy by combining two different models rather than using the same model for both generation and extraction, showing the channel is lossy and asymmetric. _Developers evaluating LLM reasoning pipelines can follow benchmark results like this one on daily.dev._ ### What causes round-trip failures when a language model converts a structured expression to text and another model tries to recover it? At least 73.6% of round-trip failures originate at the generation step, not extraction. Difficulty is driven by properties of the tree structure itself, including operator count, depth, and right-branching, rather than which model family produced or consumed the text. _Anyone tuning prompts for structured reasoning tasks can track findings like this on daily.dev._ ### Does fine-tuning on a small dataset improve open-weight language models at serializing tree-structured expressions to and from natural language? Yes, fine-tuning on roughly 3600 examples sharing the evaluation's operators and tree shapes lifts every open-weight model above untrained Gemini-3.1-Pro. A disjoint-domain test with new operators and vocabulary also improved every open-weight model, confirming the gain isn't just from matching semantics, though a gap to frontier models remains. _Teams deciding whether to fine-tune smaller models for structured reasoning can follow research like this on daily.dev._ ## Similar posts on daily.dev - [Cross-Model Experiments — The Recursion Institute](https://daily.dev/posts/cross-model-experiments-the-recursion-institute-v60yajb1s) · Medium · 2 upvotes · 0 comments - [The Machine Learning Practitioner’s Guide to Speculative Decoding](https://daily.dev/posts/the-machine-learning-practitioner-s-guide-to-speculative-decoding-3z4bou5te) · Machine Learning Mastery · 2 upvotes · 0 comments - [Your Data Model Is the Semantic Layer](https://daily.dev/posts/your-data-model-is-the-semantic-layer-ro8bktzht) · MotherDuck · 0 upvotes · 0 comments - [\[2604.20817\] Convergent Evolution: How Different Language Models Learn Similar Number Representations](https://daily.dev/posts/2604-20817-convergent-evolution-how-different-language-models-learn-similar-number-representation-krifuumbe) · Hacker News · 1 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#deep-learning](https://daily.dev/tags/deep-learning), [#nlp](https://daily.dev/tags/nlp), [#google-gemini](https://daily.dev/tags/google-gemini) [View this post on daily.dev](https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models","url":"https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf"},"datePublished":"2026-09-29T14:37:39.457Z","dateModified":"2026-09-29T19:19:22.215Z","description":"A research study introduces a round-trip protocol to measure how much tree-structured information (like arithmetic expressions) survives when language models...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Apple Machine Learning Research","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Apple Machine Learning Research","logo":"https://media.daily.dev/image/upload/s--mD-1uSe0--/f_auto,q_auto/v1785654191/logos/mlapple","url":"https://daily.dev/sources/mlapple"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,deep-learning,nlp,google-gemini","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Apple Machine Learning Research","item":"https://daily.dev/sources/mlapple"},{"@type":"ListItem","position":3,"name":"The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-communication-bottleneck-a-round-trip-study-of-tree-structured-expression-serialization-in-lang-zaagsxlyf#faq","mainEntity":[{"@type":"Question","name":"How much does accuracy vary when swapping which language model generates versus extracts structured information in a round-trip serialization task?","acceptedAnswer":{"@type":"Answer","text":"Accuracy shifts by up to 60.4 percentage points depending on which model generates the word problem and which extracts the arithmetic expression back. The best-performing pair reaches 92.9% accuracy by combining two different models rather than using the same model for both generation and extraction, showing the channel is lossy and asymmetric. Developers evaluating LLM reasoning pipelines can follow benchmark results like this one on daily.dev."}},{"@type":"Question","name":"What causes round-trip failures when a language model converts a structured expression to text and another model tries to recover it?","acceptedAnswer":{"@type":"Answer","text":"At least 73.6% of round-trip failures originate at the generation step, not extraction. Difficulty is driven by properties of the tree structure itself, including operator count, depth, and right-branching, rather than which model family produced or consumed the text. Anyone tuning prompts for structured reasoning tasks can track findings like this on daily.dev."}},{"@type":"Question","name":"Does fine-tuning on a small dataset improve open-weight language models at serializing tree-structured expressions to and from natural language?","acceptedAnswer":{"@type":"Answer","text":"Yes, fine-tuning on roughly 3600 examples sharing the evaluation's operators and tree shapes lifts every open-weight model above untrained Gemini-3.1-Pro. A disjoint-domain test with new operators and vocabulary also improved every open-weight model, confirming the gain isn't just from matching semantics, though a gap to frontier models remains. Teams deciding whether to fine-tune smaller models for structured reasoning can follow research like this on daily.dev."}}]} ```